GL3: Make gl3drawCmd_t smaller. Step 1: shader->shaderidx, char lmtexnum

It was 112 bytes, I aim for <= 40 bytes :)

we only have 17 shaders, so they can easily be indexed with an int8
instead of wasting a full pointer.

we have up to 4 lightmap textures and lmtexnum only is an index anyway,
so that too get be an int8 (char)

src/client/refresh/gl3/gl3_main.c

@@ -773,6 +773,8 @@ void GL3_Draw3DBatchesNow()
 	curState.scroll = gl3state.uni3DData.scroll;
 	curState.lightScaleForTurb = gl3state.uni3DData.lightScaleForTurb;
 
+	gl3ShaderInfo_t* shader = NULL;
+
 	// .. but at least set the identity mat so DCFlag_IsIdentityMat can be set
 	// as a sane default
 	curState.flags |= DCFlag_IsIdentityMat;
@@ -787,7 +789,11 @@ void GL3_Draw3DBatchesNow()
 		int curFlags = curState.flags;
 		qboolean updateUni3D = false;
 
-		GL3_UseProgram(cmd->shader->shaderProgram);
+		if(cmd->shaderIdx != curState.shaderIdx)
+		{
+			shader = GL3_GetDrawCmdShader(cmd);
+			GL3_UseProgram( shader->shaderProgram );
+		}
 		if(cmd->texnum != gl3state.currenttexture)
 			GL3_Bind(cmd->texnum);
 		if(cmd->lmtexnum >= 0)
@@ -836,7 +842,7 @@ void GL3_Draw3DBatchesNow()
 				lmScales[map].B = r_newrefdef.lightstyles[cmd->styles[map]].rgb[2];
 				lmScales[map].A = 1.0f;
 			}
-			UpdateLMscales(lmScales, cmd->shader);
+			UpdateLMscales(lmScales, shader);
 		}
 
 		if(flags & DCFlag_UseColor)
@@ -903,7 +909,7 @@ void GL3_Draw3DBatchesNow()
 
 static qboolean drawStateEqual(const gl3drawCmd_t* a, const gl3drawCmd_t* b)
 {
-	if( a->flags != b->flags || a->shader != b->shader
+	if( a->flags != b->flags || a->shaderIdx != b->shaderIdx
 	   || a->texnum != b->texnum || a->lmtexnum != b->lmtexnum )
 		return false;
 
@@ -952,7 +958,7 @@ GL3_BufferAndDraw3D(const gl3_3D_vtx_t* verts, int numVerts, GLenum drawMode, gl
 
 	GLushort nextVtxIdx = da_count(vtxBuf);
 	drawCmd.idxBufOffset = da_count(idxBuf);
-	assert(drawCmd.shader != NULL);
+	assert(drawCmd.shaderIdx != -1);
 
 	// translate triangle fan/strip to just triangle indices
 	if(drawMode == GL_TRIANGLE_FAN)
@@ -1068,7 +1074,7 @@ GL3_DrawBeam(entity_t *e)
 	//glDisable(GL_TEXTURE_2D);
 	drawCmd.flags |= (DCFlag_Blend | DCFlag_DisableDepthMask | DCFlag_UseColor);
 
-	drawCmd.shader = &gl3state.si3DcolorOnly;
+	GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3DcolorOnly);
 
 	drawCmd.color[0] = (LittleLong(d_8to24table[e->skinnum & 0xFF])) & 0xFF;
 	drawCmd.color[1] = (LittleLong(d_8to24table[e->skinnum & 0xFF]) >> 8) & 0xFF;
@@ -1135,13 +1141,13 @@ GL3_DrawSpriteModel(entity_t *e, gl3model_t *currentmodel)
 	if (alpha == 1.0)
 	{
 		// use shader with alpha test
-		drawCmd.shader = &gl3state.si3DspriteAlpha;
+		GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3DspriteAlpha);
 	}
 	else
 	{
 		drawCmd.flags |= DCFlag_Blend;
 
-		drawCmd.shader = &gl3state.si3Dsprite;
+		GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3Dsprite);
 	}
 
 	verts[0].texCoord[0] = 0;
@@ -1198,8 +1204,7 @@ GL3_DrawNullModel(entity_t *currententity)
 	}
 	drawCmd.flags |= DCFlag_UseColor;
 
-	drawCmd.shader = &gl3state.si3DcolorOnly;
-
+	GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3DcolorOnly);
 
 	gl3_3D_vtx_t vtxA[6] = {
 		{{0, 0, -16}, {0,0}, {0,0}},

src/client/refresh/gl3/gl3_surf.c

@@ -270,16 +270,15 @@ RenderBrushPoly(entity_t *currententity, msurface_t *fa, gl3drawCmd_t drawCmd)
 	//       and put fa->styles[] into the 3d draw vertex?
 	memcpy(drawCmd.styles, fa->styles, sizeof(fa->styles));
 	drawCmd.flags |= DCFlag_UseLmStyles;
-	static_assert(sizeof(fa->styles) == 4, "asd"); // TODO: remove
 
 	if (fa->texinfo->flags & SURF_FLOWING)
 	{
-		drawCmd.shader = &gl3state.si3DlmFlow;
+		GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3DlmFlow);
 		GL3_DrawGLFlowingPoly(fa, drawCmd);
 	}
 	else
 	{
-		drawCmd.shader = &gl3state.si3Dlm;
+		GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3Dlm);
 		GL3_DrawGLPoly(fa, drawCmd);
 	}
 
@@ -323,12 +322,12 @@ GL3_DrawAlphaSurfaces(void)
 		}
 		else if (s->texinfo->flags & SURF_FLOWING)
 		{
-			drawCmd.shader = &gl3state.si3DtransFlow;
+			GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3DtransFlow);
 			GL3_DrawGLFlowingPoly(s, drawCmd);
 		}
 		else
 		{
-			drawCmd.shader = &gl3state.si3Dtrans;
+			GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3Dtrans);
 			GL3_DrawGLPoly(s, drawCmd);
 		}
 	}
@@ -395,12 +394,12 @@ RenderLightmappedPoly(entity_t *currententity, msurface_t *surf, gl3drawCmd_t dr
 
 	if (surf->texinfo->flags & SURF_FLOWING)
 	{
-		drawCmd.shader = &gl3state.si3DlmFlow;
+		GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3DlmFlow);
 		GL3_DrawGLFlowingPoly(surf, drawCmd);
 	}
 	else
 	{
-		drawCmd.shader = &gl3state.si3Dlm;
+		GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3Dlm);
 		GL3_DrawGLPoly(surf, drawCmd);
 	}
 }

src/client/refresh/gl3/gl3_warp.c

@@ -255,7 +255,7 @@ GL3_EmitWaterPolys(msurface_t *fa, gl3drawCmd_t drawCmd)
 
 
 
-	drawCmd.shader = &gl3state.si3Dturb;
+	GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3Dturb);
 
 	for (bp = fa->polys; bp != NULL; bp = bp->next)
 	{
@@ -699,7 +699,7 @@ GL3_DrawSkyBox(void)
 	}
 	drawCmd.transModelMat = modMVmat;
 
-	drawCmd.shader = &gl3state.si3Dsky;
+	GL3_SetDrawCmdShader(&drawCmd, &gl3state.si3Dsky);
 
 	// TODO: this could all be done in one drawcall.. but.. whatever, it's <= 6 drawcalls/frame
 	//       also, they use different textures..

src/client/refresh/gl3/header/local.h

@@ -56,6 +56,8 @@
 
 #include "../../ref_shared.h"
 
+#include <stddef.h> // offsetof()
+
 #include "HandmadeMath.h"
 
 #define DG_DYNARR_ASSERT(cond, msg) \
@@ -231,7 +233,7 @@ typedef struct
 	GLuint currentShaderProgram;
 	GLuint currentUBO;
 
-	// NOTE: make sure si2D is always the first shaderInfo (or adapt GL3_ShutdownShaders())
+	// NOTE: make sure si2D is always the first shaderInfo (or adapt GL3_ShutdownShaders() and GL3_SetDrawCmdShader())
 	gl3ShaderInfo_t si2D;      // shader for rendering 2D with textures
 	gl3ShaderInfo_t si2Dtinted; // shader for rendering 2D with textures and color tinting
 	gl3ShaderInfo_t si2Dcolor; // shader for rendering 2D with flat colors
@@ -277,10 +279,10 @@ typedef struct
 // drawcommands using gl3_3D_vtx_t, for batching
 typedef struct gl3drawCmd_s {
 	hmm_mat4	transModelMat;
-	gl3ShaderInfo_t* shader;
 
 	GLuint		texnum;
-	int			lmtexnum;
+	signed char	lmtexnum;
+	signed char	shaderIdx;
 
 	float		scroll; // for gl3state.uni3DData.scroll
 	float		lightScaleForTurb; // for gl3state.uni3DData.lightScaleForTurb
@@ -320,16 +322,48 @@ GL3_CreateDrawCmd(qboolean identityTrans)
 	ret.alpha = 1.0f;
 	ret.styles[0] = 255;
 	ret.lmtexnum = -1;
+	ret.shaderIdx = -1;
 	// the other values can remain 0/NULL
 
 	return ret;
 }
 
-
-
 extern gl3config_t gl3config;
 extern gl3state_t gl3state;
 
+enum {
+	_gl3_numShaders = 1 + ((offsetof(gl3state_t, siParticle) - offsetof(gl3state_t, si2D)) / sizeof(gl3ShaderInfo_t))
+};
+
+static inline void
+GL3_SetDrawCmdShader(gl3drawCmd_t* drawCmd, const gl3ShaderInfo_t* shader)
+{
+	if(shader == NULL)
+	{
+		drawCmd->shaderIdx = -1;
+		return;
+	}
+	ptrdiff_t offset = shader - &gl3state.si2D;
+	if( offset >= 0 && offset < _gl3_numShaders)
+	{
+		drawCmd->shaderIdx = offset;
+	}
+	else
+	{
+		assert(0 && "invalid shader!");
+		drawCmd->shaderIdx = -1;
+	}
+}
+
+static inline gl3ShaderInfo_t*
+GL3_GetDrawCmdShader(const gl3drawCmd_t* drawCmd)
+{
+	unsigned shaderIdx = drawCmd->shaderIdx;
+	if(shaderIdx >= _gl3_numShaders) // because it's unsigned this also handles shaderIdx -1
+		return NULL;
+	return &gl3state.si2D + shaderIdx;
+}
+
 extern int gl3_visframecount; /* bumped when going to a new PVS */
 extern int gl3_framecount; /* used for dlight push checking */